In this paper, they discuss impulse-based simulation, and also mention possible techniques for extending these sorts of simulations to support constraint-based dynamics.

They use a optimized collision-detection system based on a heap, which allows them to find closest features in near-constant time. After finding collisions they compute the collision impulse imparted on one body by the other, and change their trajectory accordingly. 

Constraint-based systems are more appropriate when an object isn't moving, such as a box at rest on a table. Since it's silly to continually check for collisions in this case, it seems reasonable to switch over to simply constraining the box at that point on the table. And if something somes along to impart a new force on the box, well, switch back to an impulse-based system. They don't say exactly how to determine when to make this switch, but they do suggest it's possible. 

Collisions are also hard to do in a hybrid system, since it's relatively easy to establish lower bounds in impact times for an object following a coherent ballistic trajectory. Constrained objects aren't quite so easy, they say. But they suggest a hybrid approach might be possible -- again without giving any details on what this approach looks like.